Why geotechnical survey analysis by an AI system accelerates the project
The bottleneck of the survey stage is almost always processing, not drilling. A site report runs to hundreds of pages: borehole description logs, sampling records, laboratory reports for soils and water, and appendices with coordinates and absolute elevations.
Before reaching the analysis model, all of this goes through manual data transfer from reports — an engineer retypes tables, cross-checks sample IDs against appendices, and manually assigns top elevations for each layer.
That transfer is where time is lost. A single typo in a sample ID or an elevation shift throws off the entire cross-section, and finding it requires another pass through the documents.
Deadlines for preparing the conclusion stretch, not because of the scope of work, but because of cross-checks and rechecks.
A separate problem is contradictions in laboratory reports. The same sample may appear with different density and moisture values in different appendices, while the layer description in the drilling log does not match its position in the record.
Such discrepancies are caught only by manual comparison and often remain unnoticed if the report is large.
Then interpretation diverges: an aquifer appears in one cross-section but not another, layer boundaries shift by meters, and two engineers produce different models for the same site. This is no longer a matter of convenience but a risk built into the project. We examine how AI system analysis of geotechnical surveys closes this gap.
What automated parsing of reports and laboratory data delivers
The effect is measured by four numbers taken before and after implementation: archive parsing time, the share of fields extracted without an operator, the number of contradictions caught before transfer to analysis, and the completeness of source data that went into the foundation model.
Speed of parsing the report archive
- The archive is parsed in full, not selectively: scans, laboratory tables, field logs, and survey coordinates go through in a single pass — including reports the team previously never got around to.
- Preparation of a consolidated sample for a site is reduced from weeks of manual transfer to hours, and files of different formats — up to exchange datasets in AGS — are brought to a single structure, so comparison by contractor and period becomes reliable.
Reduction in data transfer errors
- The share of fields extracted without operator involvement is recorded at acceptance as a measurable metric: manual review goes to the remainder, not the entire volume.
- Contradiction control works at input — a mismatch in layer elevation between a report and a laboratory protocol, units lost during retyping, and duplicate boreholes surface before analysis, not after the conclusion is issued.
- Each value retains a link to the source page and table, so a disputed result can be checked in minutes instead of reviewing the whole report again.
The result is measurable: less manual entry, fewer returns from analysis, and more complete source data for the foundation model.
Input data formats: scans, tables, laboratory reports
Surveys for a single site arrive in several forms at once: the laboratory provides structured exports with physical and mechanical soil properties, the project archive contains scans of technical reports, and some data exists only on paper and enters the workflow after scanning.
If parsing is built for one format, the rest are entered manually — that is exactly where days are lost and discrepancies appear between the report text and the final table.
We parse each input stream in a way that produces a verifiable result: structured laboratory exports are read by columns and dictionaries, report scans are recognized with table extraction, and mixed sets are handled by a hybrid scheme — recognition plus verification against the source image.
| Processing option | When it is suitable | Limitations |
|---|---|---|
| Parsing structured exports | The laboratory provides data as files with a consistent column structure | Requires a fixed parameter dictionary; unit discrepancies must be normalized in advance |
| Processing report scans | Paper survey archive, certificates and reports only as images | Quality depends on the scan: skew, stamps over tables, and handwritten edits reduce accuracy |
| Hybrid parsing scheme | Mixed set: some data in exports, some in scans | Requires a verification reference — it is used to set the threshold below which a value goes to a specialist for review |
When each processing method is appropriate
We start with machine-readable exports — they provide the reference against which scan recognition for the same site is calibrated. We parse the paper archive with selective control: disputed cells are flagged and returned to the engineer rather than silently entering analysis.
How we implement: from data audit to handover of a working system
We build the process so that by the time the system is configured, it is clear by what criteria it is considered working. Survey archives differ from one organization to another: scans without a text layer, laboratory tables, and custom report forms by site. That is why we start not with a model but with data inventory and metrics.
Data audit and definition of quality metrics
- Audit of the survey archive: we look at what actually exists — formats, the share of scans without a text layer, recurring templates, and gaps by borehole. The output is a data map and a list of what will need to be recognized.
- Document parsing scheme: we define the extracted fields — absolute elevations, layer thickness, groundwater level, laboratory parameters — and the rules for tying them to borehole coordinates.
- Setup on customer templates: we adapt rules and dictionaries to your forms. If a section is called “Column” in one report and “Section” in another, this is taken into account explicitly rather than guessed.
- Definition of quality metrics: we fix what counts as correct field extraction and what error rate is acceptable. Without this, acceptance turns into a matter of taste.
- Verification on a holdout set: documents not involved in setup are run separately. On familiar examples, the system always looks better than it is.
- Acceptance of the result: we compare fields, review discrepancies, retrain on disputed cases, and only then hand over the system.
Verification on a holdout set and acceptance
If accuracy on critical fields has not reached the agreed level, we return to the parsing scheme rather than delivering the system with caveats. Handover includes access credentials, documentation on the rules, and training for your team — a new report can be added without us.
Delivery: parsing model, verification rules, documentation, procedures
We hand over not a “trained model” as a black box, but the entire parsing system: configurations, verification rules, and documentation that allow the output to be reproduced and checked without us.
The scope is fixed by an acceptance certificate — it is clear exactly what has been accepted into operation.
Parsing model and validation rules
- Model weights and the feature dictionary tied to a version number: rollback to the previous build takes minutes.
- Input validation rules: required fields, depth and elevation ranges, coordinate tying, and rejection of duplicate boreholes.
- Confidence thresholds at which a disputed layer goes to a geologist for review rather than being written into the report silently.
- A control export with reference markup — quality is checked on it after every update.
- Document package parsers: scans, laboratory tables, and text reports are brought to a single structure.
Documentation and update procedure
- Description of the data schema and export format to your geographic information system.
- Instructions for launch, logging, and troubleshooting for your support service.
- Model update procedure: who initiates retraining, on what data, and how the result is checked before publication.
- Versioning rules: every change to the model or rules receives a number and an entry in the change log.
Access to our engineers during the support period is included in delivery: we deploy the system together, go through acceptance and the first updates, and after that you maintain it yourselves.
Case: reducing the time to prepare a site conclusion
The customer is a design institute that prepares conclusions for industrial construction sites. One site received 40–60 survey borehole logs from different crews: with different bottomhole depths, their own soil nomenclature, and their own description in each file.
The engineer consolidated them into a single geological model in GEO5 manually, and it was precisely during consolidation that layers were lost.
Preparing the consolidated model took 10–12 working days, and discrepancies in the description of the same soils surfaced only during final proofreading.
From a sample of 200 archived conclusions, we calculated that in 6% of cases, a layer recorded in a separate borehole log was missing from the consolidated model — on site, this led to revision of foundation decisions.
What it was before and what it became after
| Metric | Before | After |
|---|---|---|
| Time to prepare the consolidated model | 10–12 working days | 2–3 days |
| Missed soil layers | 6% of conclusions | isolated cases, found automatically |
| Share of manual consolidation edits | about 70% | 15–20% |
What we changed: heterogeneous survey data are brought to a unified description schema, the model is assembled automatically, and the engineer checks not everything but only where heterogeneous borehole logs meet and disputed intervals.
Manual editing remains where geological interpretation is needed: the specialist argues not with the data but with their interpretation. Six months later, the institute moved all sites to this process, including those where surveys were conducted with a gap of several years.
How neural network scan parsing differs from ordinary recognition
Recognition answers the question “what is written,” while parsing answers the question “what does this mean for the dataset.” Optical recognition will return the string “soft-plastic loam, 4.2” as text, but it will not say which borehole it comes from, at what depth the layer was encountered, or in what units the thickness is given.
Then reconstruction of the report structure begins: the table grid, multi-level headers, merged cells, and laboratory footnotes.
Each number is tied to a borehole, a depth interval “from and to,” and a sample ID — without this, two identical records from different boreholes are indistinguishable, and the depth column loses order after simple text extraction and stops matching the descriptions.
A separate task is units of measurement. One package may contain meters and centimeters, mg/L and mg-eq/L, percentages, degrees, values such as “traces” and “not detected,” and intervals specified in words.
Values are brought to a single scale and checked for physical plausibility: layer thickness cannot exceed the depth of its base, and the sum of thicknesses cannot exceed the total borehole depth.
Only after this is geological interpretation possible: matching descriptions with drilling data and field tests, identifying engineering-geological elements, and checking their uniformity across the area.
The output is a ground model with characteristics for each layer, not a text file that an engineer parses manually.
The practical difference is the cost of an error: a missed or incorrectly tied layer enters foundation analysis and surfaces on the construction site.
Technical doubts: accuracy, formats, and quality control
We measure extraction accuracy not by eye: on a reference sample of reports, we calculate the share of correct values for each field — layer thickness, base elevation, soil category.
Each value comes with a link to the source, and discrepancies with the source are not recorded silently but flagged for manual review.
A separate difficulty is archives: scans without a text layer, lost graphics, and gaps in individual intervals. The system does not fill gaps but records them and builds a completeness map — it shows where a conclusion is based on facts and where an assumption begins.
What to do with incomplete archives?
Parsing proceeds block by block: if an interval description breaks off, the field remains empty and receives a “no data” flag instead of a substituted value.
The completeness map shows which boreholes have enough information for analysis and where additional investigation is needed.
Are mixed formats a problem?
It requires normalization, but it is solvable. Reports, scans, tables, and drawings are brought to a single schema: coordinates, elevations, layer thickness, soil classification.
Discrepancies between blocks of the same report are not averaged but shown to the engineer as a conflict for resolution.
How does the result get into analysis systems?
Export is in exchange formats and via an API, so data enters analysis without manual transfer. During repeated surveys, the new dataset is overlaid on the already loaded one, and borehole changes are visible line by line.
Who is responsible for quality?
Quality control is two-level: automatic verification of extracted values against the source and selective review by an engineering geologist.
The share of fields sent for manual review serves as a measurable acceptance criterion — it shows whether accuracy is increasing on your archives.
Let’s discuss your task: data, scope, and timeline for the first phase
Task analysis begins with your materials, not general questions. From the composition of the archive and the quality of the source data, it is already clear which areas can actually be automated and which will require manual preparation, and why.
The sooner we see real data, the more accurately we can estimate the scope and timeline of the first phase.
- What to send: two or three survey sets for different sites — borehole logs, laboratory test records, geological-lithological cross-sections, and explanatory notes.
- Formats — as is: tables, scans, drawings, and field log exports. There is no need to bring the archive into a unified form in advance; we will work with what is available.
- Archive volume: how many sets and sites, how many boreholes and pages in each, and for what periods the data was collected.
- How the material is processed now: where consolidated tables are kept, who checks discrepancies, and at which steps data is most often lost.
- How we review the data: we separately note what is already structured, what is recognized automatically, and what remains with the engineering geologist.
- First-phase options: a narrow scope on one report type, full parsing of one site, or stream processing — with different amounts of manual refinement.
- How the result is accepted: which system your specialists work in and what export formats are needed at the output.
We will come back with a plan and estimate: composition of the first phase, processing scheme, list of checks for the geologist, timelines, and points where access to historical data will be needed. Describe the task and send the materials — we will discuss the details and come back with a plan.
